Brown bears (Ursus arctos (Linnaeus, 1758)) are famous for exploiting annual pulses of Pacific salmon (Oncorhynchus sp.), but studies of how bears do so have largely been confined to anomalous locations where they congregate to feed at natural salmon migration bottlenecks. We addressed this knowledge gap using nine years of non-invasive genetic detections of brown bears along six small streams – grouped into trios along the northern and southern shores of Lake Aleknagik that constitute foraging neighborhoods where distinct bear populations prey on sockeye salmon (O. nerka (Walbaum, 1792)) – to test hypotheses about sex-specific patterns of relative stream use in relation to salmon abundance and stream morphology. Numbers detections per bear on a stream during the summer salmon spawning season were inversely proportional to water depth in the northern stream neighborhood; by contrast, greater in-stream salmon abundance boosted detection rates, and females were more likely to be detected repeatedly than males, in the southern stream neighborhood. Thus, bear behavior along small streams was consistent with optimal foraging theory, as more intensive use of streams in a season corresponded with greater abundance and accessibility (reduced stream depth) of salmon, and with the idea that female bears are less sensitive to humans, as these streams received a low but regular level of human visitation. Our findings also reveal, however, that the factors affecting bear movement patterns along small streams can vary, even among stream networks are separated by just a few km, underscoring the complexity of the bear-salmon relationship. Finally, we documented marked variability in the use of streams by individual bears, with some consistently visiting particular streams over the summer but most being detected just once, implying that impacts and reliance on salmon on any stream may differ considerably among individuals in brown bear populations.

Field data collection

We quantified use of the six focal streams by bears during the summers of 2013-2019 and 2021-2022 (no data were collected in 2020 because of the coronavirus pandemic) using a non-invasive genetic approach described in detail by Wirsing et al. (2018, 2020). Briefly, over the course of the roughly six-week salmon spawning interval, we collected hair from passing bears by deploying two barbed wires (one upstream and one downstream within the first 2 km from the mouth) crossing each stream perpendicularly at a height of ca. 50-55 cm from the ground (the optimum for snagging hair from bears stepping over or under the wire; Quinn et al. 2022). A video assessment revealed that bears encountering a wire deposited hair ~ 81% of the time (Wold et al. 2020), making this approach an effective, albeit conservative, means of documenting bear presence on a stream, at least in the vicinity of the wires. Hair samples snagged on the wires were collected regularly (usually every other day) in coin envelopes, stored in desiccant, and a subsample was sent to the Laboratory for Ecological, Evolutionary and Conservation Genetics (LEECG) at the University of Idaho for DNA analysis to assign each detection to a unique individual (see ‘Laboratory analysis’ section below). Given that each summer we deployed two hair-snagging barbed wires per stream, we could achieve up to two detections of an individual bear on a stream since the last wire check; for the purposes of this study, we defined an independent detection of an individual bear to be any instance over the course of a summer when hair collected from either wire on a given stream yielded at least one confirmed genotype. Accordingly, a bear could be detected a maximum of one time during a span between two wire checks (typically two days) on any stream. Consequently, over the ca. 6-week summer sampling season, the maximum number of detections possible for any individual bear on a stream was ca. 21.

The six focal streams have been surveyed for salmon abundance for many decades as part of the University of Washington Alaska Salmon Program, a long-term study of salmon population dynamics; patterns of salmon predation by brown bears on these streams have been investigated since the late 1980s (Quinn et al. 2001, Quinn et al. 2017). These surveys, which are performed by three people walking up the stream while counting all live and dead salmon including those in the immediate riparian zone, are prone to missing some salmon that bears move from the stream corridor (Quinn et al. 2009) but nevertheless yield a standardized, quantitative index of salmon abundance that can be used for comparisons among streams and years and linked to bear presence. They have also served as the basis for careful and repeated measurements of depth and width for each stream; notably, these values have remained stable through the years included in this study (Quinn et al. 2017, unpublished data).

Laboratory analysis

The genetic methods we used for this study were detailed previously (Wirsing et al. 2018, 2020). Briefly, DNA was extracted from snagged bear hair in a facility dedicated to low quality DNA samples using the DNeasy Blood and Tissue Kit (Qiagen, Inc.). Genotypes were generated for each sample in a multiplex polymerase chain reaction (PCR) using 10 nuclear DNA microsatellite loci and a sex locus; each sample was amplified two to four times to detect genotyping errors. Consensus genotypes were achieved following the rule that each allele needed to be observed twice per locus, and that a sample had to contain a consensus genotype at eight or more loci to be included in the matching analysis using the software Genelex (Peakall and Smouse 2006). Each sample, all from brown bears, was assigned to a specific bear along with information on its sex.

Statistical analysis     

Of the 223 individual brown bears detected over 9 years of sampling spanning an 10-year interval, 13 (5.8%) were detected in both stream neighborhoods, and of these only 3 (1.3%) switched neighborhoods (i.e., crossed the lake or outlet river) during a summer. Thus, though switching occurs, its rarity, particularly during the same summer, led us to consider the two steam trios as functionally distinct systems for the purposes of analysis. Accordingly, we modeled bear detection data from each stream neighborhood, north and south, separately. For each neighborhood, analyses were conducted using generalized linear mixed effects models (GLMM), with a negative binomial distribution built with the lme4 package in RStudio (initial Poisson models were over dispersed; Bates et al. 2015). The response variable was bear capture count, or the number of wire check (i.e., two-day) intervals during which a bear was detected on a given stream during a given summer. We included covariates for sex (female or male), mean stream depth (m; fixed across years), stream width (m; fixed across years), log-transformed in-stream salmon abundance (one estimate per stream per year), and log-transformed neighborhood salmon abundance (i.e., total salmon abundance in the other two streams of the neighborhood; one estimate per stream per year) as fixed effects in the global models for both analyses. We tested for the inclusion of random effects in negative binomial models for both neighborhoods, and likelihood-ratio tests (LRT) indicated that the treating year (χ2 (1) = 14.54, p = 0.0001) and bear ID (χ2 (1) = 63.55, p < 0.0001) as random effects improved model fit for the north neighborhood. Bear ID also improved model fit for the south neighborhood (χ2 (1) = 70.83, p < 0.0001), whereas year did not (χ2 (1) = 1.22, p = 0.27). We chose, however, to eliminate bear ID as a random effect from both analyses. The structure of the bear ID variable, given that most bears were detected once during the study, resulted in random-effect variance estimates being nearly zero (i.e., little variation across random effect levels) when added to our model sets (Bates et al. 2015, Matuschek et al. 2017). Global models for both neighborhoods thus included all predictor variables with year being treated as a fixed effect in the south and a random effect in the north. Candidate models were ranked according to differences in Akaike’s Information Criterion (∆AIC), and we considered models with a ΔAIC ≤ 2 relative to the top model to be equally supported (Burnham & Anderson, 2002). Within the set of competitive models for each neighborhood, we used the model(s) with the fewest variables as the basis for inference under the principle of parsimony (Aho et al. 2014). Covariates in these models were considered significant if the 95% confidence interval (CI) for their coefficient did not overlap zero, and odds ratios were calculated for significant covariates as a measure of effect size. For both neighborhood models, we deemed prediction 1 to be supported if a significant coefficient indicated that odds of detection on a stream increased with salmon abundance, prediction 2 to be supported if greater salmon abundance in nearby streams (i.e., neighborhood abundance) significantly reduced the odds of detection on a focal stream, prediction 3 to be supported if the odds of detection increased with reduced stream depth and/or width, and prediction 4 to be supported if females had significantly greater odds of detection on a stream than males.